<p>Fused Deposition Modelling (FDM) produces dimensional and geometric deviations that are highly sensitive to process parameters and difficult to predict using analytical models. In industrial practice, printing parameters are commonly selected through trial-and-error procedures or by minimizing overall deviations across all measurable characteristics, without considering whether these characteristics are functionally required by the engineering drawing. Such approaches often introduce unnecessary optimization constraints and may not guarantee tolerance compliance for the final part. This study proposes an ANN-based parameter selection methodology for tolerance-compliant FDM manufacturing using a drawing-driven optimization approach. An experimental database was established through a Taguchi L16 design conducted on a Zortrax M200 printer using Z-ABS material. Sixteen dimensional and geometric specifications, including dimensional errors, flatness, parallelism, perpendicularity, cylindricity, and position deviations, were measured using a Coordinate Measuring Machine (CMM). To compensate for the limited size of the experimental dataset, a physically constrained data augmentation procedure was applied by perturbing continuous process parameters within machine-resolution limits, generating 5488 training samples. Sixteen independent multilayer perceptron (MLP) models were then developed, one for each specification, and evaluated using a grouped five-fold cross-validation protocol designed to prevent data leakage between training and validation sets. All models achieved coefficients of determination (R²) greater than 0.99 with RMSE values below 0.002, demonstrating strong predictive capability within the investigated parameter space. The predictive capability of the developed models was further confirmed through an independent experimental validation campaign using five additional FDM specimens manufactured with process parameter combinations excluded from both the original design and the augmentation procedure, yielding a maximum absolute prediction error of 0.0195&#xa0;mm. The trained ANN models were then integrated into an optimization framework in which only the specifications relevant to the considered part are included. A Global Quality Index (GQI) was introduced as the average ratio between predicted deviations and allowable tolerances in order to evaluate candidate solutions. Two optimization strategies were investigated according to manufacturing constraints: a continuous optimization approach based on a Genetic Algorithm (GA), and a discrete exhaustive search over manufacturable parameter combinations. The proposed methodology was applied to a test case involving four functional specifications extracted from an engineering drawing. Both optimization strategies converged toward nearly identical parameter configurations while satisfying all tolerance requirements, with GQI values of 0.3946 and 0.4013, respectively. The optimal discrete parameter combination was further validated through physical FDM manufacturing and CMM inspection, confirming that all selected functional specifications satisfied their prescribed tolerance requirements under real manufacturing conditions. The results demonstrate that the proposed approach can effectively support tolerance-compliant parameter selection while avoiding unnecessary optimization of non-critical specifications. The methodology is directly applicable to other FDM systems once the corresponding ANN capability models are established.</p>